É«É«À²

The title of the thesis: Towards Sample- and Compute-Efficient Condition Monitoring of Induction Machines

Thesis defender: Nada El Bouharrouti
Opponent: Prof. Lucia Frosini, Universita degli Studi di Pavia, Italy
Custos: Prof. Anouar Belahcen, Aalto University School of Electrical Engineering

Induction machines are widely deployed across industrial applications due to their robustness and cost-effectiveness, yet their reliability is frequently compromised by bearing degradation and rotor-related faults. These faults often develop progressively and remain difficult to detect under limited data availability and constrained sensing conditions. By consequence, this doctoral thesis investigates sample- and compute-efficient condition monitoring strategies for fault detection in induction machines. It develops model-based, data-driven, and hybrid methods for the condition monitoring of these machines while explicitly addressing limitations in data volume and computational resources.

The research leverages analytical modelling, multi-physics simulation, feature engineering, and machine learning techniques. First, a radial lumped-parameter bearing model is introduced to capture localized fault dynamics with reduced computational complexity. This model is then extended through co-simulation with electromagnetic models to reproduce realistic rotor-stator interactions, while preserving computational efficiency. To complement this, data-driven strategies based on multi-rate signal analysis of vibration signals and transfer learning on filtered time-frequency representations are introduced, demonstrating that reliable fault diagnosis can be achieved from reduced, non-intrusive measurements. Finally, a hybrid generative framework based on generative adversarial networks is introduced to fuse simulated and experimental data for augmentation purposes. This approach produces physically consistent synthetic signals that retain key fault signatures of rotor bar failures, thereby enhancing model generalization and mitigating data scarcity.

The novel findings confirm that reliable fault diagnosis in induction machines can be achieved with reduced data and computational requirements when physical modelling and learning-based techniques are carefully considered.

Key words: Condition Monitoring, Induction Machines, Rotor Fault Detection, Ball Bearing Fault Detection, Model-Based Condition Monitoring, Data-Driven Condition Monitoring, Hybrid Condition Monitoring, Sample Efficiency.

Thesis available for public display 7 days prior to the defence at .

Contact: nada.elbouharrouti@aalto.fi